Agents

Run AI coding agents headless from swamp workflows and record what each run did and cost, so a workflow can hand simple subtasks to a cheaper model and keep the expensive one for planning and review. Supports Kiro CLI (@cachehit/agents/kiro).

Each run stores an invocation record (model, success, failure class, Kiro credits, duration, tool calls, answer preview) and a transcript with the full prompt and answer. The readonly tool profile, the default, limits the agent to reading and searching files; actor gives it every tool of a Kiro agent without permission prompts.

Quick Start

swamp extension pull @cachehit/agents
swamp model create @cachehit/agents/kiro kiro \
  --global-arg defaultModel=claude-haiku-4.5

swamp model method run kiro listModels
swamp model method run kiro invoke \
  --input prompt="Summarise what src/ does" --input cwd=/path/to/repo
swamp model method run kiro invokeAndParse \
  --input prompt='List the TODOs in src/ as JSON [{"file":..,"line":..}]' \
  --input model=claude-sonnet-4.6

Methods

• invoke: run Kiro with a prompt; fails the step if the run fails.
• invokeAndParse: like invoke, and the answer must contain JSON, which is stored as parsed; asks again when it does not.
• listModels: store Kiro's models with their credit rate multipliers.

First release with @cachehit/agents/kiro. invoke, invokeAndParse and listModels run Kiro CLI headless and record the model, credits, duration, tool calls and failure class of every run.